Azure High Performance Computing
Articles,  Blog

Azure High Performance Computing

Coming up, we explore what is now possible with high-performance computing and Microsoft Azure, including how you can use it for
enabling powerful visual simulations and engineering scenarios. Dynamic rendering for GPU intensive work, and compute scenarios for deep learning. Microsoft Mechanics Please join me in welcoming Corey
Sanders from the Azure Compute team. It’s great to be here Rick. So high performance computing, this category has kind of been around for a while now, but how’re things involving inside of
Azure to be able to support the HPC workload. Using scientific computing for
scientific analysis and computation, modeling, it’s definitely not a new thing. It’s been around all the way back to punch cards when you were sort of getting started in computers. But, over the last few years we’ve seen
a lot of massive strides in the space both from a hardware and from software perspective. And Azure is seeing the most
exciting changes as part of that. From a hardware side we now offer VM
sizes with best-in-class GPUs that I’m going to show you a little bit later. A set of sizes for computation and
a set for visualization to really be able to see things up and close. We also just announced H-Series for really fast interconnect between high-performance machines that you may want to deploy. The nice thing about this is that these unique offerings are not only single VM offerings, but they’re also multi-node for massive,
massive computation. It would blow your top off, Rick. I’m sure it would. I look forward to seeing it. But just like the rest of Azure, we’re offering both Windows and Server in Linux categories. Yeah that’s right, so with these sizes
we have built-in grid capability from Nvidia and we offer both windows for that support and for the computation side, both
windows and Linux offer this great capability. So that’s an awful lot of talk now, but how
about you go ahead and actually show it to us? Yeah, let’s take a little look. So, right here I’ve got a demo running here. If you imagine that I was an automotive engineer. And I was going to be working
with these N-series machines. If I were doing this modeling right here, I can now see the live model of these running minivans. And I know this is your car of choice,
Rick to be able to do this. And the key thing here is that with all
of these we offer discrete devices assignment, So you can get direct hardware level performance even though this is running through
a virtual machine in the cloud. And so you see I’ve got this model running,
and you can see it’s a wireframe model. If I hit play on this guy, now you can see it’s running this simulation. So from here, I’m connected to a machine in the cloud but, I can directly interrogate and see the frame and all the impacts that’s
happening with that model. And so this is utilizing of course the
GPU side for the visualization. But, prior to doing this I ran this model using CPUs to be able to calculate
what this will look like. So it’s a nice combination of GPU and CPU, again being able to scale out in the cloud. And so, what’s my job is finished I can now see my outcome, and see the results of the tests that I ran. So, typically in the olden days, you have to
go ahead and move this data all around. That’s right, so in the past when you
end up running these models you ended up you know doing
them in one location then bringing them back On-Premises
and be able to do the visualization there. Now you can sort of do it, all in the cloud. And giving you sort of another example of this. I have another demo here. And to really show you the power of that visualization, What I’ve got here is I’ve got a drill. It’s a model of a drill and as you’ll see, I’ll zoom in here. You can see I can I can zoom all the way in, and see sort of all of the details of this drill. So this is a model that’s been done with
CPU calculation building this model out. And the key thing here with what we’re
offering with our visualization skews is with grid capability from Nvidia, now I can get a very very beautiful picture
of this same exact model. And so, if I change here this is now showing you a
fully-formed model picture of this. And I can do again that same sort of zoom, but it looks so realistic. And this is thanks to the wonderful work with Nvidia to be able to offer this natively on our GPUs. Look at that, so now you zoom all the way in and you can see just really really pretty Crisp picture of that same exact model that I just showed you. So everything lives in the cloud and you
don’t have to move terabytes of data around. Scenarios and iterations around those
scenarios can happen a lot faster. That’s right, it’s incredibly fast. And it’s really easy to see this
type of model almost instantaneously. That was a great example of how you can use Azure HPC for manufacturing and engineering to work with. Let’s take a look at a different scenario, one that’s going to be a little bit
closer to my heart which is gaming. How does this help with video rendering, video effects if your designer or if you’re any kind of an artist? Yes so with this you can now
to do photorealistic rendering. and get a very very crisp picture of games or any sort of modeling you maybe doing. Movies and so on. And so what I’m showing here
now is actually a demo of a sky diver. And this is running through a client called Teradici. And so what this does it gives me a very very fast connection to this remote machine. Makes a really pretty visualization
of this remote machine, and allows me to watch this skydiver in action. And so, there is a fully modeled visualization right here. Here she came in of course by plane as any skydiver would. So you can see all of the effects in this. You can see the fog and the atmosphere impact. You can see the cloth on her outfit and it looks so real. This is really only possible when you’re using the GPU power that we enable here. She’s skydiving and there’s her pal Frank, the eagle. There they are and of course, it goes into slow motion just to really show you that impact. So I mean, it literally feels like the
server is running under her desk when she’s doing this type of model. But it isn’t, it’s in the cloud. And so, that grid capability and that quadro support natively available in the machines as
you spin them up in the cloud. Now what if I want to do any real number
crunching and data deep learning scenarios? Yeah if you’re doing scenarios where
you’re doing number crunching such as let’s say high frequency trading,
monte carlo simulations. We’ve now accelerated that compute process here made really really fast. This is the compute side of the house,
so not the visualization side. In fact many of our internal products are using this for deep learning to train things like machine learning. Being able to do things like
Skype translator or Cortana. All of these are using these types of capabilities I mean one of the cool things about this
is it’s not just a single machine. You can actually tie together thousands of nodes given infiniband connectivity between these instances under 3 microseconds between these nodes. It’s just fast. Why don’t I actually show you this in action. So what I’ve got here I’m actually
running a Linux server here. I’ve got actually two Linux servers and
I’m connecting showing both here. On the left I’ve got the CPU side. On right I’ve got a GPU side. This is kind of a classic learning
example where it’s trying to learn handwriting. And so if I start this up here I’m going
to start here on the left. And then I’m going to go over
here to the right and start this one too. And so what this is doing is going through thousands and thousands of examples of handwriting. And it’s trying to learn from human handwriting to be able to detect later on when someone writes a number, what number it is. It’s a learning algorithm going through it. And just to show you an example here, the right side is running the GPU and the left side is running CPU. Here on the left side it’s starting up and
if you see here the samples per second. It’s running at about 830 samples per second. So it’s reading 830 of these handwritten things. On the right side it’s running
12,000 samples per second. So just the difference in the scale and capability on the GPU side is pretty astounding. On the right side you can see
seventy-eight percent usage of the GPU. So it’s really pushing the full power of this machine. The left side is still kind of sluggishly moving on. So the capability to read through this many samples and learn from this algorithm this fast is just really really mind-blowing. And so as you see, it continues to run here upwards of twelve thousand here a second. And it’s going through about let’s
say about 80,000 of these so it’s going to take just a few
seconds for this to finish up. And we’ll go through all the samples and and be done. And so you can see it’s finished
now and it’s completed and fully done. All the samples here on the right are done. And the left is still kind of moving along. It’s going to take a little while. Let me show you what this actually does. So I’ve done all this learning now. Now I want to apply that learning and actually try and guess a number that’s being written. Let’s write the number 9 here, okay. So with that, let me actually go in and run the script. You can see there it validated
and it looked at the number. It read from all the learning that I’ve done in the back. And it has predicted that it has told
me this number is number nine. And so using that learning on the computation side, I’ve been able to figure out this number Ok, so that’s really cool stuff man.
Whats next? So there’s a lot of direction that we’re seeing here. The growth and the capability that we’re seeing on the computational side is just mind-blowing. Some really cool stuff is happening. As I’ve shown you today already,
a lot of new hardware shipping. And we continued to up the performance in Azure. They really are cutting edge in the
marketing offering the fastest of GPUs, the best interconnect between
instances for multiple GPUs, And of course the fastest CPU’s with the fastest interconnect there as well. We’re just seeing so many of these capabilities turn
into these intelligent services. So as I showed here, being able to learn and being able to use that learning
in real-world scenarios. Guessing a number from a handwritten is a good example of a real-world scenario. But of course, even things like Cortana intelligence. Being able to use solutions like Azure machine learning, our machine learning platform. HD insight, our Hadoop platform. With all these, you can now take
advantage of this power. Take advantage of our analytics Support and be able to get some really cool end-to-end solutions. I recommend you give it a try. So how easy is it to go ahead and try
some of these capabilities yourself? It’s amazingly easy. All you do is go on to,
sign up for an account and you can use GPUs today. Thanks a lot man. Don’t forget to keep on checking out
mechanics to show the latest in technology updates. Thanks for watching. Microsoft Mechanics

One Comment

  • Dave Voyles

    Manufacturing / Engineering
    1:55 – Car model simulation
    3:15 – Lightbox 3D model visualization

    Gaming / Visualization
    4:38 – 3DMark Sky Diver demo

    Deep Learning / ML / Compute
    5:55 – Overview, CPU vs GPU perf comparison

    8:55 – What's Next

Leave a Reply

Your email address will not be published. Required fields are marked *